Most AI assistants wait to be asked. You open a chat, paste in some context, get an answer, and the next morning you do it all again. But the useful work in a data team is rarely a one-off question. It is the Monday revenue check, the weekly churn list, the alert when a table stops updating. That work needs someone who owns it, not a box you type into.
Kin is our answer. Each kin is a named teammate with a role, a set of tools, and a memory. You give it a job in plain words, it works through the steps on its own, and it reports back to your inbox. It starts with data, because that is where we work.
What a kin actually does
You start from a template (a data analyst, a monitor, or a reporter) or write the role yourself, and choose which connections and tools it may use. Then you give it work. A task is not a single reply: the kin plans, reads the schema, writes and runs queries, pulls in a file or a web page when it needs one, and keeps going until it has an answer it can stand behind.
Here is the shape of a typical run, from an example kin watching weekly revenue:
You can watch it happen. Each step streams in as the kin works, and every task keeps its full log afterwards: each query, each page, each file, each tool call.
It checks its own numbers
The failure that worries people most with AI on business data is the confident wrong number. A model that says revenue fell 14% when the query said 4% is worse than no model at all, because someone will act on it.
So before a kin finishes, every figure in its answer is checked against what its tools actually returned, either directly or through one simple step like a difference, a sum, a ratio, or a percent change. The check is plain code, not another model call. A figure with no backing goes back to the kin once to fix. If it still cannot be backed, you see it flagged rather than reading it as fact.
It asks before it acts
Reading is cheap to get wrong. Writing is not. Anything that changes the world outside the task waits for you: changing data in a source, sending an email to someone other than you, calling a webhook, posting to Slack. You get an approval card with exactly what the kin wants to do, and you can approve it, edit it, or reject it.
- A kin can only write to sources you have explicitly marked writable
- If you edit an action before approving it, the kin is told what changed
- You decide per kin which actions, if any, may run without asking
It works on its own schedule
A teammate that only works when you remember to ask is not much of a teammate. A kin can wake up four ways:
- On a schedule: every weekday at 8, the first of the month, whatever cron can express
- On a webhook: from your own systems, when something happens
- On an email: forward it something and it gets to work
- On a data change: when a table it watches changes
Results and alerts land in your Kin inbox, with a link back to the full task.
It remembers, and gets better
Kin keeps two kinds of memory. Facts and preferences (that internal test accounts should be excluded, that "region" means the billing region) are remembered across tasks, and you can review them at any time. Runs that went well become playbooks the kin can reuse next time, so a job it has done before does not start from scratch.
What it can reach
Kin works with the databases and warehouses the ThinkingLanguage platform connects to, from Postgres and MySQL to Snowflake and BigQuery, plus files you upload (CSV, spreadsheets, PDFs), the public web, and any MCP server you choose to connect. For heavier data work it runs ThinkingLanguage, our open-source language where tables are a native type, so analysis happens in code that runs, not in the model's head.
Try Kin
Free for 3 kin and 300 tasks a month. Kin is invitation only while in early access.
Request an invitation Meet KinWhere it runs, and what we do with your data
Kin is hosted on our cloud, so there is nothing to install. That also means it is for data that is allowed to be in the cloud. If yours is not, Foundry and Studio are the products built to run inside your perimeter.
Inside Kin, connection credentials are encrypted at rest and never shown to the model. Web access goes through a sandbox that blocks private and internal addresses. Accounts get breached-password checks and optional two-factor sign-in. Your data is processed to run your tasks and is not used for training, as the privacy policy sets out.
Pricing
Every plan uses the default Kin model with usage included, or you can bring your own API key for OpenAI, Anthropic, Groq, or any OpenAI-compatible endpoint.
- Free: 3 kin, 300 tasks and 1M model tokens a month
- Pro, $49 a month: 25 kin, 5,000 tasks and 20M model tokens a month, and it also covers the hosted ThinkingLanguage platform
- Custom: more capacity and organization workspaces, talk to us
We are onboarding teams in small batches so we can support each one properly. If you have a job you would hand to a kin tomorrow, tell us what it is, and we will send you an invitation.
// give it a job, not a prompt
← back to blog